Research On Feature-based Semantic Role Labeling For English And Chinese

Correct and automatic semantic parsing has always been one of major goals in natural language understanding. With the solid development of syntactic parsing during the past decade, semantic parsing has been drawing more and more attention recently. However, due to difficulty in deep semantic parsing, current researches focus on shallow semantic parsing, which attempts to label predicate-related constituents in a sentence with semantic roles, such as agent and patient, and has been widely applied in information extraction, question and answering, machine translation etc.This paper will tackle semantic role labeling, a representative shallow semantic parsing. Although recent years have seen much progress in semantic role labeling, there exists some key issues. First, the performance of current semantic role labeling systems heavily depends on the performance of syntactic parsing, especially in Chinese. Second, it is hard to further improve the performance of semantic role labeling. To address above issues, this paper systematically explores feature-based methods in semantic role labeling. We focus on:1) Constituent-based SRL. In particular, various structural features are investigated and optimized, a head-driven pruning algorithm is proposed and two post-processing mechanisms are explored, to further improve the performance. Experimental results show that our system achieves best-reported performance for SRL on the top-best parse tree.2) Dependency relation-based SRL. This is done by porting the above constituent-based SRL system to the dependency tree structure, with focus on the development of specific features and improved pruning algorithms in capturing dependency relations. In addition, systematic analysis and comparison are performed between constituent and dependency relation-based SRL. Experimental results that our system achieves best-reported performance on the gold parse tree.3). SRL in Chinese. This is done by porting the above two SRL systems in English to the Chinese language, with focus on exploring specific characteristics in Chinese. Experimental results on constituent-based SRL show that our system achieves best-reported performance in Chinese. In addition, we pioneer the work of building a fully automatic dependency relation-based SRL and predicate detection platform in Chinese.The contributions of this paper lie in systematic and in-depth research on semantic role labeling in both English and Chinese from both the constituent structure and the dependency tree structure, e.g. the head-driven pruning algorithm and the post-processing mechanism. Our research significantly improves the performance of SRL and thus exhibits an important reference value to the future work in semantic parsing.